Adaptive Control for Autonomous Mobility via LoGenE: Reward-guided Genetic Evolution of LoRA Adapters

Abstract

Autonomous docking in mobile robots requires precise control in dynamic environments that are affected by sensor noise and surface variations. Although PID controllers are widely used because of their simplicity, fixed gains often fail to adapt to environmental variability. Recent studies have explored large language models (LLMs) for dynamic PID tuning; however, their high computational overhead limits real-time deployment. To address this issue, we propose LoRA-based genetic evolution (LoGenE), a gradient-free neuroevolution framework that optimizes lightweight low-rank adaptation (LoRA) adapters for dynamic PID control. LogenE evolves lightweight adapter modules offline using control logs, eliminating the need for gradient updates or expensive real-time simulations. The resulting models are deployable on devices with minimal latency. Experiments conducted in the ROS + Gazebo simulation environment showed that LoGenE significantly improved docking performance compared to a base LLM, demonstrating robust and adaptive control suitable for real-world robotic systems.